Geoscience Reference
In-Depth Information
9.4.5
Simulation of Switzerland
One of the near-future goals of the performance improvements presented in this
chapter is to perform simulation runs for the whole of Switzerland; therefore,
a first experiment in this direction is conducted. This experiment simulates the
whole population of Switzerland (7.3 million agents) on a network with around
one million links. The agents traveled using different transportation modes, such as
car, bus, bike, and on foot. The experiment is run with the single-threaded version
of JDEQSim and parallel event handling with a single thread. It took around 3 h
and 16 min for a single iteration of micro-simulation and event handling, while
for replanning and the rest, an additional 70-80 min are needed. As MATSim is
an iterative process, depending on the search space, many iterations are needed to
reach a relaxed state (Balmer et al. 2009 ). When only route choice, mode choice,
and departure time/duration adaption are enabled, around 60 iterations are required
(based on experience). It is estimated that it may take around 11 days for these
60 iterations to complete, considering that the overhead of writing out events to
the hard drive is only conducted each 10th iteration. If we assume a speedup of
four for JDEQSim compared to JQueueSim and also take the computational time of
the other modules into account, it is estimated that with JQueueSim it would take
around 36 days to calculate such a scenario. This is a speedup of around 3.2 for the
overall simulation. While the performance gains achieved by the work presented in
this chapter are important, this means that more progress related to performance is
still needed, which is discussed in the next section .
9.5
Discussion and Future Work
While two ways to significantly shorten the runtime of the MATSim simulation are
presented in this chapter, it seems like the “low-hanging fruits have been picked”
and additional improvements will not be as straightforward and may possibly lead
to not as much performance improvement, as well as requiring major changes, to
the existing models and interfaces. In the following section, issues involved are
discussed together with possible solutions.
9.5.1
Parallel Micro-simulation
In order to achieve a major breakthrough with regards to the micro-simulation
performance, making use of multiple threads seems to be crucial. Although a
first success in the direction of a parallelization of DEQSim has been made, a
successful integration of this into MATSim needs more work and would also require
fundamental changes to the existing models.
Two points seem central to a successful integration of parallel JDEQSim in
MATSim: decoupling of threads and integration of parallel JDEQSim with parallel
Search WWH ::




Custom Search